Research map / Video / Video segmentation and tracking
Event-assisted Low-Light Video Object Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes an event-assisted end-to-end framework for low-light video object segmentation, with adaptive cross-modal fusion and event-guided memory matching, plus synthetic LLE-DAVIS and real-world LLE-VOS datasets.
Problem
Video object segmentation degrades under low-light conditions because image quality drops and query-memory matching becomes inaccurate; no event-based low-light VOS dataset existed.
Approach
Fuses image and event features with learned adaptive filters (ACMF) and uses event and memory-mask priors to gate key features before matching (EGMM) within an AOT-style matching VOS framework.
Key results
- On the synthetic LLE-DAVIS set, the method reaches J&F 0.628 versus 0.559 for AOT (about +6.9 points).
- On the real-world LLE-VOS set, overall J&F is 0.678 versus 0.633 for DeAOT and 0.607 for Zero-DCE+AOT.
- Ablation on LLE-DAVIS: baseline J&F 0.584 rises to 0.601 with ACMF and 0.628 with ACMF+EGMM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add event input to image input (ablation) | VOS framework · LLE-DAVIS | J &F: 0.559 -> 0.628 | Helped |
| use event input only instead of image input (ablation) | VOS framework · LLE-DAVIS | J &F: 0.559 -> 0.547 | Hurt |
| add ACMF to feature concatenation (ablation) | VOS framework · LLE-DAVIS | J &F: 0.584 -> 0.601 | Helped |
| add EGMM after ACMF (ablation) | VOS framework · LLE-DAVIS | J &F: 0.601 -> 0.628 | Helped |
| use mask input only instead of no EGMM prior (ablation) | EGMM · LLE-DAVIS | J &F: 0.601 -> 0.464 | Hurt |
| use event input only instead of no EGMM prior (ablation) | EGMM · LLE-DAVIS | J &F: 0.601 -> 0.552 | Hurt |
| use both event and mask inputs instead of no EGMM prior (ablation) | EGMM · LLE-DAVIS | J &F: 0.601 -> 0.628 | Helped |
| use 1 EGMM block instead of 3 (ablation) | VOS framework · LLE-DAVIS | J &F: 0.628 -> 0.603 | Hurt |
| use 2 EGMM blocks instead of 3 (ablation) | VOS framework · LLE-DAVIS | J &F: 0.628 -> 0.618 | Hurt |
| use 4 EGMM blocks instead of 3 (ablation) | VOS framework · LLE-DAVIS | J &F: 0.628 -> 0.622 | Hurt |
Surprising findings
- Zero-DCE low-light enhancement as a preprocessing step does not consistently help VOS; on LLE-VOS, Zero-DCE+AOT underperforms plain AOT on some metrics and DeAOT plus Zero-DCE falls below DeAOT on indoor J&F.
Limitations
- Real-world dataset is small (70 videos), and the test split has only 20 clips.
- Models are trained from scratch on the low-light datasets, so the comparison does not use large pretrained VOS weights.
- Event-only input is weaker than image input in ablation, so the method depends on image quality being partially usable.
Experimental setup
Datasets: DAVIS 2017 LLE-DAVIS LLE-VOS
Benchmarks: LLE-DAVIS LLE-VOS
Metrics: J F J&F
Baselines: STCN XMem AOT DeAOT Zero-DCE+STCN Zero-DCE+XMem Zero-DCE+AOT Zero-DCE+DeAOT
Builds on: AOT ID module (identity assignment) ESIM Zero-DCE
Closest papers by meaning
- Efficient Event-Based Semantic Segmentation via Exploiting Frame-Event Fusion: A Hybrid Neural Network Approach (AAAI 2025 · similarity 0.88)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.87)
- ESEG: Event-Based Segmentation Boosted by Explicit Edge-Semantic Guidance (AAAI 2025 · similarity 0.86)
- Know Where You Are From: Event-Based Segmentation via Spatio-Temporal Propagation (AAAI 2025 · similarity 0.86)
- FlexEvent: Towards Flexible Event-Frame Object Detection at Varying Operational Frequencies (NeurIPS 2025 · similarity 0.86)
- EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.